Looking for breakthrough ideas for innovation challenges? Try Patsnap Eureka!

Feature space reformation-based teacher-student architecture noise-tag-containing image classification method

A technology of feature space and classification method, applied in neural architecture, neural learning method, biological neural network model, etc., can solve the problem of ignoring the benefits of mutual learning, to promote self-learning, improve performance, robustness and effectiveness excellent effect

Pending Publication Date: 2022-03-11
匀熵教育科技(无锡)有限公司
View PDF0 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

These methods exploit the differences between models to avoid confirmation bias, but ignore the benefit of mutual learning

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Feature space reformation-based teacher-student architecture noise-tag-containing image classification method
  • Feature space reformation-based teacher-student architecture noise-tag-containing image classification method
  • Feature space reformation-based teacher-student architecture noise-tag-containing image classification method

Examples

Experimental program
Comparison scheme
Effect test

Embodiment Construction

[0027] 1. Related work

[0028] 1.1 Noise Label Learning

[0029]Due to its strong learning ability, deep neural networks can easily adapt to any noise. Training a deep neural network on a noisy data set allows the network to fit noise samples, which greatly reduces the generalization ability of the model. To overcome this problem, some researchers modify the loss function to make the model more robust to noise. This approach has strong theoretical support, but as the complexity of the noise increases, the effectiveness of the approach gradually decreases. At the same time, the modification of the loss function increases the amount of computation required for training convergence. Since learning noisy labels directly cannot achieve good results, researchers instead consider correcting noisy labels. Reed et al. [Reed, Scott, et al. "Training deep neural networks on noisy labels with bootstrapping." arXiv preprint arXiv:1412.6596 (2014).] combine raw labels with model predic...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention discloses a teacher-student architecture noise-label-containing image classification method based on feature space reforming, and belongs to the field of noise label learning and image classification. The success of deep learning mainly depends on a large-scale tagged data set. Most of data sets are manually labeled, so that the labeling accuracy is difficult to guarantee. Therefore, noise tags are inevitable in a real-world data set. Due to the strong learning ability of the deep neural network, the noise label is very easy to over-fit, and the model performance is reduced. In order to solve the problem, the invention provides a teacher-student structure based on feature space reformation, and the teacher-student structure is used for learning noise labels. A teacher model is trained to correct noise labels for student models. A feature space reforming mechanism is applied to guide a student model to carry out self-learning, so that the student model learns more meaningful information. We assess our methods on the three reference data sets. Experimental results verify the effectiveness of the proposed architecture.

Description

technical field [0001] The invention belongs to the field of noise label learning and image classification, and in particular relates to a method for classifying images with noise labels based on a teacher-student structure based on feature space renormalization. Background technique [0002] Deep learning has achieved remarkable success in many areas of computer vision, but its success is largely due to huge labeled datasets such as ImageNet. However, collecting such a large labeled dataset usually takes a lot of time and effort. Nowadays, most datasets are manually labeled, which leads to a problem that the ratio of noisy labels increases as the amount of data increases. Therefore, the accuracy of the labels in the dataset cannot be guaranteed. Studies have shown that deep neural networks are prone to overfitting noisy labels, resulting in weakened model performance. Therefore, how to train robust models in noisy environments has aroused great interest of researchers. ...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
IPC IPC(8): G06V10/764G06K9/62G06N3/04G06N3/08
CPCG06N3/08G06N3/045G06F18/24
Inventor 孙俊
Owner 匀熵教育科技(无锡)有限公司
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic, Popular Technical Reports.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap|About US| Contact US: help@patsnap.com